A sEMG-based gesture recognition framework for cross-time tasks
Xingguo Zhang, Tengfei Li, Yue Zhang, Maoxun Sun, Cheng Zhang, Jie Zhou · Measurement Science and Technology · 2024
Abstract In the practical application of action pattern recognition based on surface electromyography (sEMG) signals, the electrode displacement and the time-varying characteristics of the signals during cross-time signal acquisition can reduce the classification accuracy. This study designs a 12 d forearm sEMG signal cross-time acquisition experiment, introduces time span into the dataset, and proposes a cross-time gesture recognition framework based on deep convolutional neural networks (CNN) with sEMG signals. In the cross-validation of single-day analysis, recognition rates using multiple CNN modules exceed 90%. However, the average recognition rate for cross-day analysis is only 59.0%. The classification performance of the framework is significantly improved in the multi-day analysis by gradually increasing the number of training days. In particular, 97.4% accuracy is achieved in the cross-time recognition task by using a specific configuration of DenseNet as the network module and extracting features with one-dimensional (1D) convolution on signal fragments. Compared to the method of extracting short-time Fourier transform image features as input using two-dimensional convolution, the training method of extracting signal features using 1D convolution reduces the time consumed to about 1%, which is advantageous in terms of model performance.